What one GPU costs per hour
Chips are 40–55% of what a token costs to produce. This ledger records daily per-GPU hourly prices on public rental platforms: RunPod secure and community cloud list prices, Vast.ai on-demand market minimum and median (hosts with ≥95% reliability only), and the starting price on suanli.cn. Every number carries a fetch time and a raw snapshot. Most Chinese platforms show prices only after login; being connected.
| GPU | RunPod Secure cloud | RunPod Community cloud | Vast.ai On-demand min | Vast.ai On-demand median | Suanli Starting price |
|---|---|---|---|---|---|
| RTX 4090 24 GB VRAM | $0.740 ¥4.99 | $0.340 ¥2.29 | $0.136 ¥0.918 · 49 hosts | $0.469 ¥3.16 · 49 hosts | $0.249 ¥1.68 |
| RTX 5090 32 GB VRAM | $0.990 ¥6.67 | $0.690 ¥4.65 | $0.256 ¥1.72 · 52 hosts | $0.537 ¥3.62 · 52 hosts | — |
| RTX A6000 48 GB VRAM | $0.530 ¥3.57 | $0.330 ¥2.22 | $0.281 ¥1.90 · 8 hosts | $0.469 ¥3.16 · 8 hosts | — |
| L40S 48 GB VRAM | $1.09 ¥7.34 | $0.790 ¥5.32 | $0.509 ¥3.43 · 7 hosts | $0.794 ¥5.35 · 7 hosts | — |
| A100 SXM4 80 GB VRAM | $1.59 ¥10.71 | $1.39 ¥9.36 | $0.402 ¥2.71 · 18 hosts | $0.715 ¥4.82 · 18 hosts | — |
| H100 SXM 80 GB VRAM | $3.49 ¥23.51 | $2.69 ¥18.12 | $1.85 ¥12.43 · 8 hosts | $3.00 ¥20.22 · 8 hosts | — |
| H100 NVL 94 GB VRAM | $3.19 ¥21.49 | $2.59 ¥17.45 | $2.20 ¥14.84 · 4 hosts | $2.72 ¥18.29 · 4 hosts | — |
| H200 141 GB VRAM | $4.59 ¥30.92 | $3.59 ¥24.19 | $1.98 ¥13.31 · 6 hosts | $3.64 ¥24.50 · 6 hosts | — |
| B200 180 GB VRAM | $6.79 ¥45.74 | $5.98 ¥40.29 | $7.75 ¥52.23 · 3 hosts | $8.14 ¥54.81 · 3 hosts | — |
What this ledger is for
Read it alongside the Token Price Index : one side is what tokens sell for, the other is what the machines that make them rent for. Our "below cost floor" band for relay quotes draws its floor from public channel prices and rental prices like these; the conversion assumptions will be spelled out in Definitions. Here we only record, we do not speculate.
Clear direction, grounded judgement.